Manifold-Aware Retractions Overview

• Manifold-aware retractions are smooth maps that approximate the Riemannian exponential map while preserving the tangent space structure of a manifold.
• They guarantee first-order or higher-order agreement with the manifold's intrinsic geometry, making them suitable for optimization, numerical integration, and interpolation.
• These retractions enable robust algorithmic design in applications such as dynamical low-rank approximation, generative modeling, and stochastic simulation on complex manifolds.

A manifold-aware retraction is a map that approximates the Riemannian exponential map in the context of smooth or geometric optimization, interpolation, simulation, and generative modeling on manifolds and constrained sets. Retractions provide a computationally tractable mechanism for moving along prescribed tangent directions while ensuring iterates remain on the manifold. The term “manifold-aware” emphasizes that retractions are designed to respect the intrinsic geometry, ensuring at least first-order agreement with the tangent space structure at each reference point. Retractions play an essential role in manifold-based numerical integration, dynamical low-rank approximation, high-order interpolation, constrained optimization, diffusion-based generative planning, and the analysis of algorithms on both smooth and singular spaces.

1. Definitions and Fundamental Properties

A retraction $R$ on a smooth manifold $M \subseteq \mathbb{E}$ (with $\mathbb{E}$ a Euclidean space) is a smooth mapping

$R: TM \to M, \quad (x, v) \mapsto R_x(v)$

defined on a neighborhood of the zero section such that: - $R_x(0) = x$ - $D R_x(0)[v] = v$ for all $v \in T_xM$.

This construction guarantees that for small increments, the retraction curve $\tau \mapsto R_x(\tau v)$ with $\sigma(0)=x,\, \sigma'(0)=v$ makes $R_x(v)$ an order-$1$ or first-order move respecting the tangent geometry of $M$ (Séguin et al., 2023, Séguin et al., 2022, Olikier, 2024).

If, in addition, $R_x(\tau v)$ exhibits vanishing acceleration at the origin, i.e., $\sigma''(0) = 0$, then $R$ is termed a second-order retraction. Higher-order retractions further match higher derivatives of the exponential map (Séguin et al., 2023, Gawlik et al., 2017, Séguin et al., 2022). For closed subsets or singular varieties, both strong and weak retraction notions are used. The weak (Hosseini–Uschmajew) definition requires only

$\lim_{t \to 0^+} \frac{R(x, tv) - (x + tv)}{t} = 0$

with no continuity away from zero, providing sufficient structure for optimization and numerical integration (Olikier, 2024).

2. Canonical Constructions and Examples

Manifold-aware retractions have a variety of explicit realizations:

• Matrix Manifolds:
  • Fixed-Rank Matrix Manifold: Several retractions exist, including rank-$r$ truncated SVD (metric-projection), QR-based retractions, and the KLS retraction. The KLS retraction performs substeps on factor matrices (QRs), then updates the core component, resulting in a second-order retraction that is robust even for small singular values (Séguin et al., 2023).
  • Stiefel Manifold: Retractions based on polar decomposition, QR factorization, Cayley transforms, and the new “polar-light” retraction, which uniquely offers both second-order accuracy and a closed-form inverse (Jensen et al., 23 Feb 2026). For a base point $X$, and tangent $\xi$, the polar-light retraction and its inverse are given by explicit formulas with complexity $O(np^2+p^3)$ and well-characterized domains.
• Submanifolds and Embedded Varieties:
  • Projective Retraction: $R(x,v) = \text{Proj}_M(x+v)$, where $\text{Proj}_M$ denotes metric projection onto the manifold; smooth near regular points (Olikier, 2024, Séguin et al., 2022).
  • Level Set Retraction: For $M = F^{-1}(0)$ and regular $F$, $R(x,v) = (x+v) - [\nabla F(x)]^+ F(x+v)$, a (potentially analytic) strong retraction (Olikier, 2024).
  • Convex and Algebraic Sets: For closed convex or algebraic sets, the projective retraction is always a weak retraction.
• Data-driven and Algorithmic Retractions:
  • LoMAP: In generative diffusion planning, LoMAP approximates the local tangent space by PCA on nearest neighbors and retracts candidate points by projecting onto this low-rank subspace, provably enforcing manifold adherence in high dimension (Lee et al., 1 Jun 2025).
• Random Walks and Sub-Riemannian Geometry:
  • Retractions designed to match local geodesic structure (e.g., via affine connections or approximated exponential map) ensure that stochastic walks on sub-Riemannian manifolds converge, in law, to horizontal Brownian motion if the retraction is second-order (Herrmann et al., 2023).

3. Theoretical Guarantees and Hierarchy

Manifold-aware retractions must satisfy order-specific local properties. First-order retractions guarantee infinitesimal tangent alignment, while second- or higher-order retractions further match geodesic acceleration. For applications such as random walk generator convergence or high-order ODE integrators, second-order retractions are essential (ensuring, e.g., $O(h^3)$ local errors and $O(h^2)$ global error in integrators like AFE and PRH) (Séguin et al., 2023, Herrmann et al., 2023).

On closed or singular sets, only weak retractions may exist, yet this suffices for descent-based optimization: all required stationarity and line-search conditions follow from first-order agreement. Strong retractions are necessary only for higher regularity or symmetry demands (e.g., symmetry in ODE integration) (Olikier, 2024).

Manifold/Set Type	Retraction Order	Existence (strong/weak)	Inverse Available
Smooth manifold	1, 2, ≥3	Strong	Sometimes
Singular/algebraic set	1 (usually)	Weak	Rare
Stiefel manifold	1, 2	Strong (polar-light, QR)	Yes (polar-light)
Data manifold (LoMAP)	1 (PCA-based)	Weak ($C$ by data)	Trivial (proj.)

4. Computational Methods and Algorithmic Use

Retractions are central in manifold-aware algorithms:

• Optimization: Retracted line-search proceeds by stepping along the tangent direction $v$ using $R(x, t v)$. Weak retractions suffice, provided first-order agreement, for establishing convergence and achieving descent in constrained optimization (Armijo rule, quasi-Newton updates on $M$) (Olikier, 2024).
• Numerical Integration: Retraction-based integrators for ODEs on matrix manifolds leverage second-order retractions for higher local accuracy. The PRH and AFE methods use explicit formulas for higher-order updates, with PRH including Hermite interpolation and demanding efficient inverse retractions (Séguin et al., 2023). Classical projector-splitting (KSL/KLS) and SVD-based step schemes fit within this retraction framework.
• Interpolation: Hermite interpolation on manifolds employs retraction-convex sets and inverse retractions for de Casteljau-style constructions; order-$4$ uniform error is achieved under sufficient smoothness (Séguin et al., 2022).
• Random Walks: In sub-Riemannian settings, horizontal retractions approximate normal geodesics and are applied in random walk generators for the simulation of horizontal Brownian motions; second-order agreement is essential for weak convergence (Herrmann et al., 2023).
• Generative Modeling: LoMAP, in diffusion planning, retracts each sampling step onto an empirically determined local tangent subspace, decreasing infeasibility and improving statistical metrics (up to $30$–$50\%$ reduction in artifacts per (Lee et al., 1 Jun 2025)).

5. Analytical and Computational Trade-offs

The construction and selection of retractions are shaped by manifold type, application, and computational constraints:

• Order vs. Cost: Higher-order retractions typically require more complex operations (matrix exponentials, SVDs, small-scale lyapunov solves). For the Stiefel manifold, the polar-light retraction uniquely combines second-order accuracy and a closed-form inverse at $O(n p^2 + p^3)$ cost (Jensen et al., 23 Feb 2026). QR-based retractions are cheaper but only first-order; SVD-based approaches may not offer smoothness (no strong retraction on rank-deficient variety) (Séguin et al., 2023).
• Domain and Validity: Retractions are valid only locally; step sizes must remain within neighborhoods where the construction is well-posed (e.g., within the injectivity radius of the log map for polar-light retraction).
• Availability of Inverse: Many computations (notably Hermite interpolation, Riemannian barycenters) benefit from closed-form inverses, motivating the development of retractions like polar-light and Q-factor retraction on Stiefel (Séguin et al., 2022, Jensen et al., 23 Feb 2026).
• Robustness: KLS retraction for fixed-rank matrices maintains stability in the presence of small singular values, outperforming traditional SVD-projection retractions in certain dynamical low-rank settings (Séguin et al., 2023).
• Data-driven Consistency: In high-dimensional diffusion planning, the locally PCA-based retraction of LoMAP delivers rigorous error control relative to the manifold’s tangent subspace (scaling with PCA error $\epsilon$) and prevents divergence off the data manifold (Lee et al., 1 Jun 2025).

6. Applications and Empirical Performance

Manifold-aware retractions have widespread and growing applicability:

• Dynamical Low-Rank Approximation: Retraction-based integrators systematically unify many classical and novel DLRA schemes. The KLS retraction yields stable, efficient schemes robust to modeling error and numerical instability (Séguin et al., 2023).
• Optimization in Deep Learning and Geometry: Retractions are foundational in algorithms projecting onto the Stiefel or orthogonal groups for kernel orthogonalization, robust principal component analysis, and structure-preserving learning.
• Trajectory Generation and Control: LoMAP enhances feasibility and sample quality in diffusion models, with empirical results indicating improvements in return and realism scores and dramatic reductions in infeasibility metrics in challenging offline RL and planning problems (Lee et al., 1 Jun 2025).
• Stochastic Simulation: Random walk generators employing horizontal retractions yield accurate weak convergence to target sub-Riemannian diffusions, crucial in geometric probability and sub-Riemannian statistics (Herrmann et al., 2023).
• Interpolation and Data Approximation: Hermite interpolation schemes leveraging manifold-aware retractions achieve high accuracy for trajectory estimation, temporal smoothing, and data fitting on matrix manifolds and beyond (Séguin et al., 2022).

7. Topological and Geometric Context

Beyond algorithmics, retractions appear in the purely topological study of manifolds. Every topological $n$-manifold is a Euclidean neighborhood retract (ENR): there exists a continuous map $r:U\to\iota(M)$ retracting a neighborhood $U$ of an embedded manifold $\iota(M)\subset\mathbb{R}^{2n+1}$ onto the manifold (Floris, 2022). In the smooth category, this is refined via tubular neighborhoods and exponential maps (yielding smooth retractions with local convexity), while in the topological category, only continuity is required.

This generality underpins the utility of manifold-aware retractions: they are combinatorially and analytically flexible tools, linking topological, geometric, algorithmic, and data-driven strategies for handling the intrinsic constraints and structure of high-dimensional spaces.